We recently celebrated the inauguration of CIMC, with talks, panel discussions, explaining our mission and connecting with friends; recordings of the talks are out now!
RESEARCH
-
Sakana AI New Model Sparks Reinforcement Learning Revolution
By
–
Sakana AI New Model Sparks a RL Revolution https://
youtu.be/2mezj14pCFI?si
=HozvSHRv-57IC6T-
… via @YouTube -
Lidar Cars and Tesla FSD Gap to Level 4 Autonomy
By
–
Lidar equipped cars can be mass produced for $30k. And Tesla FSD is still L2, needs 100-1000x higher city miles per critical disengagement to get to L4.
-
MedGemma: Google’s Vision-Language Model for Healthcare Applications
By
–
🚀Meet MedGemma – Google’s powerful Vision-Language Model for healthcare, built on the Gemma 3 architecture.
— Satya Mallick (@LearnOpenCV) 24 juin 2025
In our article, we break down MedGemma’s core DNA, showcase real-world applications, walk you through the code, and highlight its stellar performance.
Whether you're… pic.twitter.com/Fp4hzU2dp1Meet MedGemma – Google’s powerful Vision-Language Model for healthcare, built on the Gemma 3 architecture.
In our article, we break down MedGemma’s core DNA, showcase real-world applications, walk you through the code, and highlight its stellar performance. Whether you're -
Chain of Debate: Multiple LLMs Collaborating and Discussing
By
–
Chain of Thought was just the beginning. Next up: Chain of Debate.
We’re going from a single model “thinking out loud” to multiple models discussing out loud.
Debating, debugging, deliberating.
AI becomes AIs.
“Two heads are better than one” is true for LLMs too. -
Performance gains from optimized implementation in AI systems
By
–
i'll have to page @lo_LB_La for engineering details but iirc most of the performance gain comes from the optimized implementation rather than architectural details
-
Gaussian Mixture EM Assumptions: Separation Requirement Analysis
By
–
After careful reading I saw that this analysis still requires the Gaussian mixtures to be separated for EM to succeed- something that the method of moments doesn’t require. Do you think that assumption can be removed?
-

Testing Seedance-1-Pro Physics and 9:16 Native Video Support
By
–
Time to test the physics of seedance-1-pro I guess.
— fofr (@fofrAI) 24 juin 2025
It can also do 9:16 natively. https://t.co/btkk388qEu pic.twitter.com/46z7SH5mb0Time to test the physics of seedance-1-pro I guess.
It can also do 9:16 natively. -
Diminishing Returns in Language Model Training Efficiency
By
–
the returns have been diminishing for a while, that’s certainly true NeoBERT is 250M params but trained on 2T tokens. objectively a crazy thing to do
-

New Paper on Training Great AI Models Now Available on ArXiv
By
–
and our paper is available on ArXiv now, soon-to-be-published in TMLR. if you're interested — take a read! a lot of great nitty-gritty detail in what it takes to train a great model. http://
arxiv.org/abs/2502.19587